Towards Recommender System Optimization. Part 2: How Can Artists Influence Recommendation Algorithms?

This article is the second part in our series on Recommender System Optimization (think SEO for music recommenders), aiming to present a framework for algorithmic optimization on DSPs (and Spotify, specifically). You can find the first part laying out the conceptual foundations of RSO and showcasing our algorithmic data tool over here. This piece can stand on its own — but If you haven't gone through part one just yet, we suggest you check it out first and then return to this article.

Disclaimer: we're navigating uncharted waters here. Algorithmic optimization in music is still in its infancy. The tactics mentioned in this text are yet to be tested at scale. The goal of this piece is not to offer you a bulletproof guide to algorithmic optimization but to showcase various levers you could use to influence streaming recommenders. If you're interested in exploring this developing field and working with us on the algorithmic optimization of assets you manage, you can apply to our beta consultancy program here.

So, how can artists and their teams influence recommender systems? First, let's ask ourselves: for a given track on Spotify, what inputs would a recommender process to decide how to recommend it? 

There are two decisions the algorithm will make that ultimately determine the track's success:

• The recommender will qualify the track to work out where it should be recommended i.e., generate a track representation, allocating the track to a set of candidate pools. For example, a track from Artist A would be part of the candidate tracks for the radio playlist of Artist B, as they share sonic and user affinity similarities.
• Second, it will review how listeners on Spotify interact with the track to understand when and to whom it should be recommended (i.e., assess user feedback and determine track ranking for each respective candidate pool). 

Both of these decisions are based primarily on the track itself, with the artist's past release activity playing a supporting role. For example, if the artist has already released ten lo-fi hip-hop tracks, their new single is more likely to be qualified within that genre. Or, if the artist is historically well-received by streaming audiences, the newly released single will get a boost to its algorithmic rankings. 

So, recommender system optimization is a two-fold problem:

• How can we provide optimal inputs that will "pull" towards the target algorithmic positioning and get the track on the high-traffic, priority candidate pools? 
• How can we drive qualified, engaged, and algorithmically relevant audiences to the track (through both paid and organic channels) to validate these initial connections and improve your track's ranking within these pools?

But what concrete actions can we take to "optimize the inputs"? Perhaps, it's best to try to break down the problem into smaller bits first. To help structure the RSO workflow, let's look at its closest counterpart: search engine optimization. A typical SEO operation would act along the three main avenues: technical SEO, on-page SEO, and off-page SEO. A similar structure can be applied to RSO:

• Technical RSO: optimizing the technical distribution pipeline to ensure that the artist catalog and metadata attached are complete and accurate.   
• On-platform RSO: optimizing all signals accessed by the recommender directly on streaming platforms 
• Off-platform RSO: optimizing all "external" signals, originating outside streaming platforms but ultimately influencing the artist's algorithmic streaming performance.

Technical RSO: Optimizing metadata inputs & assessing your distribution pipeline 

This first technical aspect of RSO is a basic layer of algorithmic optimization you can already find implemented by record labels/distributors. Just like technical SEO, seeking to make the website's contents & page metadata accessible to search engines, technical RSO aims to ensure that the recommender system is supplied with all the data it needs to qualify artist releases.

The first aspect of technical optimization is all about the capabilities and overall quality of your distribution pipeline. Recommender systems use a wide range of artist-supplied data to generate asset recommendations: from genre and mood tags to songwriter credits, release dates, and lyrics. This is the first step of technical optimization — ensuring that there's a direct, bug-free stream of metadata from artist teams to DSPs and that all the relevant data passed down from artist teams (e.g., producer and songwriter credits) is accurate and complete. 

Besides, asset structure errors like compound or duplicate artist pages are still quite common — and I guess I don't have to tell you that having one of your key tracks uploaded under a separate artist profile won't score any extra points with the recommender. So, a technical audit of the artist profile is also a must if you want to ensure that your RSO operation stands on a solid foundation. 

The second aspect of the technical RSO is all about "subjective" metadata fed into the recommender through artist-facing interfaces (like the S4A pitch form). Which genre, style, and mood tags are optimal for the upcoming release? Today, there is rarely a definitive answer to questions like that. That said, with the right insights in hand and clear algorithmic targets — you can use our algorithmic mapping to help inform that decision — the metadata input can become the most straightforward and efficient way of influencing the recommender. After all, that's precisely why Spotify for Artist pitch form has all the genre and style fields: to get direct artist input into the recommender to help it pre-qualify the release while there's still little internal data to educate the system.

Finally, the last component of the technical RSO is concerned with platform-specific features and the overall artist profile quality. Leveraging Spotify-specific features like merch shelves, canvas, artist pick, or synced lyrics is a sure way to signal to the recommender that the project is an active, commercial priority. This, in turn, will likely boost the ranking of all assets released under that project — and so ensuring that the artist page on Spotify is complete becomes another technical RSO priority.

Yet, technical optimization is hardly the most exciting or impactful subset of RSO. Technical RSO is a must-have foundation of your algorithmic strategy — but the real value is what you can build on top of it.

On-platform RSO: Optimizing product features, catalog structure, and playlisting strategy

The second essential building block of an RSO strategy is on-platform optimization, drawing parallels to on-site SEO — a practice of optimizing elements accessed by the search engine directly through the website (as opposed to optimizing external signals, collectively known as "off-site SEO"). Similarly, on-platform RSO is concerned with optimizing signals that can be accessed by the recommender directly on Spotify.

Optimizing Product Features

After the release is uploaded on the platform, Spotify will run raw audio analysis algorithms against the recording to generate the song's audio profile. It's hard to say what the output of that analysis will look like, as it remains a part of the secret recipe for each platform. Yet, an educated assumption would be that state-of-art recommenders can generate highly detailed audio profiles for any given track — not only extracting the primary audio features (danceability, energy, valence, etc.) but also assessing the song's structure, melodic composition, and more. To get an approximation of the track's audio profile, we can employ third-party audio analysis solutions, like Cyanite.ai:

‍

‍

The second aspect of the recording analyzed is the lyrics — if there are any, of course. The recommender will analyze the text with Natural Language Processing algorithms to establish the general theme and meaning of the lyrics and extract any valuable keywords, such as brands, locations, people, or other cultural phenomena mentioned. So, the track's lyrics should also be considered with the assumption that the recommender will have a pretty robust idea of the song's meaning.

Now, there are a few ways how you could go about integrating such insights into the RSO strategy. For most companies and artists, the creative vision comes first — meaning that the audio/lyrics analysis should be applied once the song is finished. In that case, the insights can be used to guide the algorithmic positioning of the track, helping the RSO manager determine which recommendation context would fit the track. For instance, with "Glimpse of Us" it would make sense to shoot for "sad" and "romantic" recommendation settings — frosting algorithmic connections with artists/songs with similar "sad" and "romantic" mood profiles within Joji's algorithmic surroundings, adjusting the playlisting strategy to target "breakup" playlists, and so on.

In other cases, however, your target algorithmic positioning can guide the production process instead. Aiming to produce a perfect soundtrack for the morning run? Pass an early demo through audio analysis software and check the energy levels — or maybe even set up a benchmark by analyzing tracks on Spotify editorial playlists fitting your target recommendation context. 

These two approaches — designing the track to fit your target algorithmic positioning vs. building your positioning around the audio/lyrical profile of a track that's already finished — could be treated as opposite sides of the spectrum, leaving plenty of room for hybrid approaches. The audio analysis could come in at the latter stages of the production process (e.g., mixing and mastering) as an additional input that would help guide creative decisions rather than dictating them. But perhaps it's best not to dwell too much on the timeless "business vs. art" dilemma — so let's move on to the next part.

Optimizing Release and Collaboration Strategies

To get to this second building block of on-platform RSO, we have to zoom out from single-release optimization to a broader level of optimizing artist catalog structures. To draw parallels to on-site SEO once more: it's rarely enough to publish a single article addressing a keyword to amplify a website's search performance. Even if that article is really, really good in fulfilling the target search intent. Instead, SEO managers usually have to produce a series of articles to cover a specific topic relevant to their target audience to signal the website's topical expertise to the search engine and build domain authority. Similarly, if you're looking to optimize algorithmic performance on DSPs, it's rarely enough to optimize just a single release.

Having spent the last year studying artists' algorithmic profiles, we've learned that streaming algorithmic audiences are rarely homogeneous. In Joji's case, for instance, our model has extracted eight distinct artist clusters, each corresponding to a subset of Joji's algorithmic audience: 

‍

One made up of listeners of chill R&B, another — bringing in fans of more energetic hip-hop of BROCKHAMPTON and JPEGMAFIA, a third one — leaning towards moody bedroom pop of boy pablo and Phum Viphurit. The list could go on, but the point is that to manage such fragmented algorithmic communities, RSO strategy should aim to develop a catalog that would address each of these sub-audiences in its own right. Or, at least, address algorithmic sub-audiences considered long-term priorities by the artist team.

Perhaps, one of the singles of the next release cycle should come with a remix package to bridge the gap to the world of electronic music. Maybe another could feature a metal-inspired beat that would appeal to audiences looking for a more vibrant sound. Collaborations allow you to create direct algorithmic links, becoming an extremely versatile and efficient instrument for enriching algorithmic profiles. And when we say collaborations, it's not limited to releases featuring other performing artists. Most streaming services have access to song credits data, and Spotify is the most advanced DSP in leveraging songwriter data. So, the track's songwriter and producer data will also be fed into the recommender, helping establish connections between songs and projects with shared contributors. 

So, which performing artists, songwriters, and producers should an artist collaborate with? Are they compatible with their algorithmic profile? Will they help strengthen the link to existing algorithmic sub-audiences or foster new algorithmic connections that align with your RSO targets? If the project had a sharp change in sound with the latest release cycle, how could we bridge the "algorithmic gap" between the new and the old sound to ensure a smooth algorithmic transition? These are just some of the questions involved in on-platform RSO, aiming to develop a catalog that will perform well in algorithmic spaces.

Optimizing Playlisting Strategy

The final component of on-platform RSO is the playlisting strategy. It's no secret that third-party and editorial playlists play a massive role in on-platform promotion — while user-generated playlists often become the primary source of long-term retention. However, if we adopt RSO optics, playlists take up another critical role — composing the data set used by the recommender to establish similarity between tracks on the platform. We've already reviewed playlist-based collaborative filtering in great detail in our breakdown of Spotify recommender, so let's keep it broad strokes. 

The logic is quite simple: if a person puts two songs on the same playlists, these tracks are, in some way or another, similar to each other. So, once two tracks get playlisted together enough times, the recommender will treat it as a clear signal of similarity. This means that playlists featuring the track — or, rather, other tracks playlisted along your release — play a crucial role in shaping the track's asset representation.

An educated assumption would be that the dataset used to train the recommender is mainly made up of user-generated, "personal" playlists. However, editorial playlists still play a major role in determining the release playlist profile: if the track is exposed to thousands (if not millions) of listeners on New Music Friday, the chances are it would make it onto a lot of user-generated playlists — alongside other tracks off this week's edition of NMF. In that sense, the larger the playlist following is, the more significant its impact on the track's similarity profile, even though the playlist itself might not provide direct input into the recommender system.

On-platform RSO raises an additional set of new criteria to conventional playlist promotion. If you're looking to build a robust algorithmic profile for a track, your playlisting strategy should set goals beyond maximizing streams. It's imperative that the playlists you pitch to contribute positively to your target algorithmic positioning — i.e., feature other artists and tracks that will be a relevant similarity input.

In that sense, less is more sometimes — rather than spreading wide and trying to get on any playlist out there, it might be wise to hyper-target your playlisting campaigns to a more narrow but algorithmically relevant subset of playlists reflecting your song's mood, style, and genre. And I guess it goes without saying that getting on low-engagement, bot-ridden playlists is not among RSO's best practices. Tracklists of such playlists would usually be composed of a random selection of paid placements, which will foster redundant algorithmic connections and confuse the recommender. Besides, a poor listening experience on low-quality playlists would surely translate into low engagement from any real listener that tunes into the playlist — which would further penalize your track's ranking. 

‍

Off-platform RSO: Optimizing Marketing Strategy, from PR and Word-of-Mouth to Paid Advertising and Artist Communities. 

Finally, we get to the last component of RSO strategy — off-platform RSO. Just like off-page SEO deals with all "external" aspects of search optimization (backlink and traffic generation, for instance), off-platform RSO deals with all signals influencing the track's performance that originate outside streaming platforms.

Optimizing Press Coverage & Online Word-of-Mouth.

In a way, this first aspect of off-platform RSO is very similar to the SEO backlink strategy. Except, it's not the links that we're after — but all online mentions of our artists and their tracks. To qualify artists and tracks on the platform, Spotify crawls online music media websites and prominent discussion platforms (i.e., Reddit, Twitter, and Facebook), applying NLP models to extract adjectives, genre tags, and artist references used to describe artists and their releases. This data is then passed into the recommender system as a final component of the dataset used to generate asset representations.

In this sense, RSO opens up a new frontier for music PR. As current economic turmoil hits the industry, we see music businesses slashing down their PR budgets to prioritize marketing activities that directly impact streaming volume and short-term revenues. This new RSO angle of PR argues against that trend: apart from building up long-term brand value and general awareness, a well-executed PR campaign educated by RSO analysis can impact the artist's algorithmic performance. That said, there's still a certain scale requirement — the waves you make need to be loud enough to be heard by the recommender, and getting on a few local music blogs won't get you there.

Your PR strategy (and the brand you build around the artist in general) is a powerful driver of the artist's algorithmic profile. Online mentions play a vital role in defining asset representations, and sometimes a simple RSO-driven edit of the press release (i.e., mentioning target artist references, moods, and genres) can come a long way in amplifying your algorithmic potential. This RSO-centric role of PR becomes exceptionally crucial for emerging artists with no historic online coverage to speak of. If you're starting from scratch, it's much easier to enrich the artist's online coverage with RSO-appropriate references and tags — turning PR strategy into one of the most potent "algorithmic levers" for the artist team when optimizing asset representations. 

Yet, that also means it's just as easy to steer the artist profile in the wrong direction. Sometimes, the artist's breakout story can generate a decent amount of organic press & online clout — picture a TikTok-fueled viral success story, for example. But, if the PR isn't addressed at the right moment, the artist might be categorized by the recommender as "TikTok-core" based on all the organic "TikTok trend coverage" — damaging their algorithmic potential if they were to try and break away from that initial launchpad. So, in that sense, press coverage is a bit of a double-edged sword. Which is why controlling the narrative around the artist is a must for any emerging project looking to leverage algorithmic recommendations strategically.

Optimizing Paid Traffic: Digital Advertising & Other Promotion Channels

With online coverage completing the list of inputs influencing artist representations on Spotify, we are finally in the asset ranking optimization territory. By way of a refresher, an asset rank is a function of user feedback, reflecting how well the song performs in a particular consumption context on Spotify. When it comes to ad-driven listenership — listeners coming to DSPs through a promoted link to their profile, asset, or playlist — the context should roughly fall under "active discovery". A user is logging onto the platform to interact with a specific asset: in that setting, the recommender is likely to focus on discovery engagement metrics, meaning save rates, playlist additions, follows, and downstream consumption (i.e., if the user digs further into the artist catalog/playlist instead of interacting with just a single track).

From that perspective, RSO is generally aligned with digital advertising. In the current streaming economy, establishing any sort of meaningful ROI on advertising budgets generally means that a single ad click must generate multiple streams down the line — turning saves, playlist additions, and catalog exposure into crucial KPIs for digital marketers. Yet, there's a critical distinction between the two. When it comes to conventional advertising practices, the listening profile of your traffic is pretty much inconsequential: a stream is a stream, and a new fan is a new fan. If we apply RSO optics, however, the listening profile of your target audience becomes very important.

Imagine a track with only three algorithmic connections, meaning that its candidate pool profile consists of just three other songs: tracks A, B, and C, all produced by different artists. Of these three tracks, only track A is considered a priority by the RSO manager, as it's more popular than the others (and thus likely to bring more algorithmic traffic). To elevate the asset ranking, the label launches an incentivized "listen to unlock" advertising campaign on Instagram, driving engaged traffic to the release page on Spotify. The track scores well in saves and playlist additions, and the campaign is an apparent success — yet, the algorithmic traffic that follows is still underwhelming. 

The reason? Well, these listeners that we drove to Spotify, did they listen to tracks A, B, or C in the past, or at least other tracks of the same artists? Or did the campaign reach listeners of different genres and artists not currently algorithmically related to the advertised track? It's not enough to drive highly engaged listeners to the track to improve asset ranking. Instead, you need to drive highly engaged listeners exposed to the relevant artists behind the candidate pools you're trying to rank up on. In other words, to rank up in a candidate pool for BLACKPINK, you need to generate positive user feedback from K-pop fans, cementing the notion that "this artist is similar to BLACKPINK, and some of BLACKPINK's fans love it — so it should be recommended to more BLACKPINK fans."

From the RSO perspective, marketing campaigns should focus on driving engaged, qualified traffic while paying close attention to what other artists and tracks these listeners are likely to listen to. In that sense, RSO is once again calling for hyper-targeting of marketing efforts: a single new listener that would help you rank up on target candidate pools might be more valuable than 10 "random" ad-driven streams. 

In some cases, a significant promotion event can also profoundly affect asset representations. Take syncs, for example — an impactful placement is likely to send waves across the streaming platforms and the Internet as a whole. The following web-based discussions around the show, playlists featuring the show's soundtrack, and resulting listening activity are likely to shift the representation of the release. For instance, the algorithmic radio playlist for Kavinsky's Nightcall still features Under Your Spell by Desire — simply because both tracks got syncs in the movie Drive some 11 years ago.

And, to wrap up this section on driven traffic: don't buy streams. I can't stress this enough — amassing fake streams is a guaranteed way of ruining your track's algorithmic profile. Ironically, artists commonly engage with "bot farms" with the opposite intention, like trying to pass arbitrary algorithmic thresholds of "20k streams to get on Discover Weekly". In our year of experimentation with Spotify recommender, we found no proof of such algorithmic traffic thresholds. I mean, sure, additional listenership does help the recommender qualify and rank the releases, but ask yourself: how would the recommender treat a track for which half of the streams are 31 seconds long, with zero saves, shares, or clicks to profile? 

Optimizing Organic Listenership

However, for most artists with developed audiences, such "driven" traffic will be just a fraction of the overall traffic directed to streaming platforms. A vast share of streaming traffic will be generated by re-engaging existing audiences across social media and other platforms — and so the reception of the track by the artist's existing audience plays a critical role in defining its algorithmic rankings. In this context, just like with driven traffic, the RSO manager has to pay close attention to how engaging the track is with existing fans and what other artists and tracks this core fanbase is likely to listen to.

Do fans obsess over the new track, play it on repeat, share it with their friends, and save it to the library or personal playlists? Or do they listen to it once to never come back? A track that over-indexes in engagement with the artist's core fanbase is likely to improve its algorithmic rankings across the board, so it's essential to generate as many saves, playlist additions, and repeat listens as possible from existing fans. That said, driving audiences to DSPs is a key goal of most social media strategies, and it's probably not our place to educate you on how to run an Instagram profile. 

However, from the RSO perspective, you have to consider your core audience's algorithmic inertia. When the artist's fanbase makes it to Spotify, what are its effects on the algorithmic profile of the track? Let's assume that the track performs really well with your core audience — how would that affect its algorithmic potential? Essentially, there are two ways you can approach this challenge of algorithmic inertia: 

• Study your fans' listening habits: learn what other artists and tracks your fans are listening to the most, and integrate that knowledge into your RSO strategy. Understand where your current audience is likely to pull the track to assess the feasibility of algorithmic targets.
• Influence your fans' listening habits: under this second, more proactive approach, you could try to sway your fans' listening habits, funneling algorithmic inertia to drive the artist profile closer to your RSO targets.

Getting Started with Your Own RSO Strategy

Well, that should cover it — we've now gone through each distinct subset of algorithmic optimization, from the technical foundations of RSO down to playlisting, PR, and marketing strategies. So, to sum up our two-part RSO series thus far, to get started with your own RSO strategy:

1. Extract and analyze the artist's current algorithmic profile (you can apply to our closed beta program here)
2. Define your algorithmic strategy, set targets, and prioritize candidate pools. Locate algorithmic spaces that are aligned with your release strategy, and aim for artists and tracks within these spaces that will bring you the most relevant traffic. 
3. Develop and implement RSO tactics following the three pillars we've covered in this text. Align different aspects of the release from PR to marketing to playlisting to pull towards your algorithmic goals.

To guide your RSO journey, we are currently developing an online course on Music Tomorrow Academy that will dive much deeper into the various RSO tactics with actionable tips, data tools, and checklists you could use to develop your artists' algorithmic profiles. Pre-enroll now through Music Tomorrow Academy to be the first to access the course once it's out!